World-in-Context 5B matches 22B EchoWM on WBench navigation

On the WBench benchmark, a 5-billion-parameter model called World-in-Context posts a navigation score of 81.5, which the authors describe as 'on par with 22B EchoWM (81.5 vs. 81.6).' The same 5B model 'attains a 93.8% perspective-switching success rate, substantially outperforming existing 5B baselines (18.8%),' according to a preprint posted to arxiv by Harold Haodong Chen and co-authors.

The authors argue existing interactive video world models bolt controls on through 'fragmented architectural pathways that capture interdependencies only indirectly.' Their alternative puts view, action and camera controls into the same self-attention space as the rest of the tokens. The paper reports that 'explicit view tokens disentangle conflicting dynamics and unlock emergent view switching from single-view data,' while action and camera inputs are described as governing 'local responsiveness and global stability' respectively. The method also adds an Interaction-Transport RoPE scheme that threads temporal attention through accumulated control transitions.

Two researchers we track posted the preprint link within hours. The abstract gives one benchmark and no inference-cost or training-compute figures.